Master Recherche IAC TC2: Apprentissage Statistique & …sebag/Slides/M2R_TC2_1_cours.pdf · Master...

Master Recherche IACTC2: Apprentissage Statistique & Optimisation

Alexandre Allauzen − Anne Auger − Michele SebagLIMSI − LRI

Sept. 16th, 2013

Where we are

Ast. series Pierre de Rosette

Maths.

World

Data / Principles

Naturalphenomenons

Modelling

Human−relatedphenomenons

You are here

CommonSense

Where we are

Sc. data

Maths.

World

Data / Principles

Naturalphenomenons

Modelling

Human−relatedphenomenons

You are here

CommonSense

Harnessing Big Data

Watson (IBM) defeats human champions at the quiz game Jeopardy (Feb. 11)

i 1 2 3 4 5 6 7 81000i kilo mega giga tera peta exa zetta yotta bytes

I Google: 24 petabytes/dayI Facebook: 10 terabytes/day; Twitter: 7 terabytes/dayI Large Hadron Collider: 40 terabytes/seconds

Machine Learning and Optimization

Machine Learning

World → instance xi →Oracle↓yi

Optimization

ML and Optimization

I ML is an optimization problem: find the best model

I Smart optimization requires learning about the optimizationlandscape

Types of Machine Learning problems

WORLD − DATA − USER

Observations

UnderstandCode

UnsupervisedLEARNING

+ Target

PredictClassification/Regression

SupervisedLEARNING

+ Rewards

DecideAction Policy/Strategy

ReinforcementLEARNING

The module

1. Introduction. Decision trees. Validation.

2. Neural Nets

3. Statistics

4. Learning from sequences

5. Unsupervised learning

6. Representation changes

7. Bayesian learning

8. Optimisation

Pointers

I Slides of this module:http://tao.lri.fr/tiki-index.php?page=Courseshttp://www.limsi.fr/Individu/allauzen/wiki/index.php/

I Andrew Ng courseshttp://ai.stanford.edu/∼ang/courses.html

I PASCAL videoshttp://videolectures.net/pascal/

I Tutorials NIPS Neuro Information Processing Systemshttp://nips.cc/Conferences/2006/Media/

I About ML/DMhttp://hunch.net/

Today

1. Part 1. Generalities

2. Part 2. Decision trees

3. Part 3. Validation

Examples

I Vision

I Control

I Netflix

I Spam

I Playing Go

I Google

http://ai.stanford.edu/∼ang/courses.html

Reading cheques

LeCun et al. 1990

MNIST: The drosophila of ML

Classification

Detecting faces

The 2005-2012 Visual Object Challenges

A. Zisserman, C. Williams, M. Everingham, L. v.d. Gool

The supervised learning setting

Input: set of (x, y)

I An instance x e.g. set of pixels, x ∈ IRD

I A label y in {1,−1} or {1, . . . ,K} or IR

Pattern recognition

I Classification Does the image contain the targetconcept ?

h : { Images} 7→ {1,−1}

I Detection Does the pixel belong to the img of targetconcept?

h : { Pixels in an image} 7→ {1,−1}

I SegmentationFind contours of all instances of target concept in image

The 2005 Darpa Challenge

Thrun, Burgard and Fox 2005

Autonomous vehicle Stanley − Terrains

The Darpa challenge and the AI agenda

What remains to be done Thrun 2005

I Reasoning 10%

I Dialogue 60%

I Perception 90%

Robots

Ng, Russell, Veloso, Abbeel, Peters, Schaal, ...

Reinforcement learning Classification

Robots, 2

Toussaint et al. 2010

(a) Factor graph modelling the variable interactions

(b) Behaviour of the 39-DOF Humanoid:Reaching goal under Balance and Collision constraints

Bayesian Inference for Motion Control and Planning

Go as AI Challenge

Gelly Wang 07; Teytaud et al. 2008-2011

Reinforcement Learning, Monte-Carlo Tree Search

Energy policy

ClaimMany problems can be phrased as optimization in front of theuncertainty.

Adversarial setting 2 two-player gameuniform setting a single player game

Management of energy stocks under uncertainty

States and Decisions

States

I Amount of stock (60 nuclear, 20 hydro.)

I Varying: price, weather alea or archive

I Decision: release water from one reservoir to another

I Assessment: meet the demand, otherwise buy energy

Reservoir 1

Reservoir2

Reservoir 3

Reservoir 4

Lost water

PLANT

NUCLEAR PLANT

DEMAND

PRICE

Netflix Challenge 2007-2008

Collaborative Filtering

Collaborative filtering

Input

I A set of users nu, ca 500,000

I A set of movies nm, ca 18,000

I A nm × nu matrix: person, movie, ratingVery sparse matrix: less than 1% filled...

Output

I Filling the matrix !

Criterion

I (relative) mean square error

I ranking error

Spam − Phishing − Scam

Classification, Outlier detection

The power of big data

I Now-casting outbreak of flu

I Public relations >> Advertizing

Mc Luhan and Google

We shape our tools and afterwards our tools shape usMarshall McLuhan, 1964

First time ever a tool is observed to modify human cognition thatfast.

Sparrow et al., Science 2011

Types of application

Domain But : Modelling

Physical phenomenons analysis & controlmanufacturing, experimental sciences, numerical engineering

Vision, speech, robotics..

Social phenomenons + privacyHealth, Insurance, Banks ...

Individual phenomenons + dynamicsConsumer Relationship Management, User Modelling

Social networks, games...

PASCAL : http://pascallin2.ecs.soton.ac.uk/

Banks, Telecom, CRN

Ex: KDD 2009 − Orange

1. Churn

2. Appetency

3. Up-selling

Objectives

1. Ads. efficiency

2. Less fraud

Health, bio-informatics

Ex: Risk factors

1. Cardio-vascular diseases

2. Carcinogenic Molecules

3. Obesity genes ...

Objectives

1. Diagnostic

2. Personalized care

3. Identification

Scientific Social Network

Questions

1. Who does what ?2. Good conferences ?3. Hot/emerging topics ?4. Is Mr Q. Lee same as Mr Quoc N. Lee ?

[tr. Jiawei Han, 2010]

e-Science, Design

Numerical Engineering

I Codes

I Computationally heavy

I Expertise demanding

Fusion based on inertial confinement, ICF

e-Science, Design (2)

Objectives

I Approximate answer

I .. in tenth of seconds

I Speed up the design cycle

I Optimal design More is Different

Autonomous robotics

Complexe, monde ferme simple, randomDesign

[tr. Hod Lipson, 2010]

Autonomous robotics, 2

Reality Gap

I Design in silico (simulator)

I Run the controller on the robot (in vivo)

I Does not work !

Closing the reality Gap

1. Simulator-based design

2. On-board trials safe environnement

3. Log the data, update the simulator

4. Goto 1

Active learning Co-evolution[tr. Hod Lipson, 2010]

Overview

Examples

Introduction to Supervised Machine Learning

Decision trees

Empirical validationPerformance indicatorsEstimating an indicator

Types of Machine Learning problems

WORLD − DATA − USER

Observations

UnderstandCode

UnsupervisedLEARNING

+ Target

PredictClassification/Regression

SupervisedLEARNING

+ Rewards

DecidePolicy

ReinforcementLEARNING

Data

Example

I row : example/ case

I column : feature/variable/ attribute

I attribute : class/label

Instance space XI Propositionnal :X ≡ IRd

I Structured :sequential,spatio-temporal,relational.

aminoacid

Data / Applications

I Propositionnal data 80% des applis.

I Spatio-temporal data alarms, mines, accidents

I Relationnal data chemistry, biology

I Semi-structured data text, Web

I Multi-media images, music, movies,..

Difficulty factors

Quality of data / of representation

− Noise; missing data

+ Relevant attributes Feature extraction

− Structured data: spatio-temporal, relational, text, videos,..

Data distribution

+ Independants, identically distributed examples

− Other: robotics; data streams; heterogeneous data

Prior knowledge

+ Goals, interestingness criteria

+ Constraints on target hypotheses

Difficulty factors, 2

Learning criterion

+ Convex optimization problem

↘ Complexity : n, nlogn, n2 Scalability

− Combinatorial optimization

H. Simon, 1958:In complex real-world situations, optimization becomesapproximate optimization since the description of the real-world isradically simplified until reduced to a degree of complication thatthe decision maker can handle.Satisficing seeks simplification in a somewhat different direction,retaining more of the detail of the real-world situation, but settlingfor a satisfactory, rather than approximate-best, decision.

Learning criteria, 2

The user’s criteria

I Relevance, causality,

I INTELLIGIBILITY

I Simplicity

I Stability

I Interactive processing, visualisation

I ... Preference learning

Difficulty factors, 3

Crossing the chasm

I No killer algorithm

I Little expertise about algorithm selection

How to assess an algorithm

I Consistency

When number n of examples goes to infinityand target concept h∗ is in H

h∗ is found:

limn→∞hn = h∗

I Speed of convergence

||h∗ − hn|| = O(1/n),O(1/√

n),O(1/ ln n)

Context

Disciplines et criteres

I Data bases, Data MiningScalability

I Statistics, data analysisPredefined models

I Machine learningPrior knowledge; complex data/hypotheses

I Optimisationwell / ill posed problems

I Computer Human InteractionNo final solution: a process

I High performance computingDistributed processing; safety

Supervised Learning, notationsContext

World → Instance xi →Oracle↓yi

INPUT ∼ P(x, y)

E = {(xi , yi ), xi ∈ X , yi ∈ Y, i = 1 . . . n}HYPOTHESIS SPACE

H h : X 7→ YLOSS FUNCTION

` : Y × Y 7→ IR

OUTPUTh∗ = arg max{score(h), h ∈ H}

Classification and criteriaSupervised learning

I Y = True/False classificationI Y = {1, . . . k} multi-class discriminationI Y = IR regression

Generalization Error

Err(h) = E [`(y , h(x))] =

∫`(y , h(x))dP(x , y)

Empirical Error

Erre(h) =1

n

n∑i=1

`(yi , h(xi ))

Bound structural risk

Err(h) < Erre(h) + F(n, d(H))

d(H) = Vapnik Cervonenkis dimension of H, see later

The Bias-Variance Trade-off

Biais Bias (H): error of the best hypothesis h∗ de H

Variance Variance of hn as a function of E

h*

h

h

Variance

h

H

Bias

target concept

Function Space

Overfitting

Test error

Training error

Complexity of H

Key notions

I The main issue regarding supervised learning is overfitting.

I How to tackle overfitting:I Before learning: use a sound criterion regularizationI After learning: cross-validation Case studies

Summary

I Learning is a search problem

I What is the space ? What are the navigation operators ?

Hypothesis Spaces

Logical Spaces

Concept ←∨∧

Literal,Condition

I Conditions = [color = blue]; [age < 18]

I Condition f : X 7→ {True,False}I Find: disjunction of conjunctions of conditions

I Ex: (unions of) rectangles of the 2D-planeX .

Hypothesis Spaces

Numerical Spaces

Concept = (h() > 0)

I h(x) = polynomial, neural network, . . .

I h : X 7→ IR

I Find: (structure and) parameters of h

Hypothesis Space H

Logical Space

I h covers one example x iff h(x) = True.

I H is structured by a partial order relation

h ≺ h′ iff ∀x , h(x)→ h′(x)

Numerical Space HI h(x) is a real value (more or less far from 0)

I we can define `(h(x), y)

I H is structured by a partial order relation

h ≺ h′ iff E [`(h(x), y)] < E [`(h′(x), y)]

Hypothesis Space H / Navigation

H navigation operators

Version Space Logical spec / genDecision Trees Logical specialisation

Neural Networks Numerical gradientSupport Vector Machines Numerical quadratic opt.

Ensemble Methods − adaptation E

Overview

Examples


Decision trees


Decision Trees

C4.5 (Quinlan 86)

I Among the most widelyused algorithms

I EasyI to understandI to implelementI to useI and cheap in CPU time

I J48, Weka, SciKit

NORMAL

>= 55 < 55

Age

Smoker

no yes

Sport

RISK

NORMAL

highlow

RISK

Tension

yesno

Diabete

yes

RISK PATH.

no

Decision Trees

Decision Trees (2)

Procedure DecisionTree(E)

1. Assume E = {(xi , yi )ni=1, xi ∈ IRD , yi ∈ {0, 1}}

• If E single-class (i.e., ∀i , j ∈ [1, n]; yi = yj), return• If n too small (i.e., < threshold), return• Else, find the most informative attribute att

2. Forall value val of att• Set Eval = E ∩ [att = val ].• Call DecisionTree(Eval)

Criterion: information gain

p = Pr(Class = 1|att = val)I ([att = val ]) = −p log p − (1− p) log (1− p)

I (att) =∑

i Pr(att = vali ).I ([att = vali ])

Decision Trees (3)

Contingency TableQuantity of Information (QI)

p

QI

0.1 0.3 0.5 0.7 0.9

0.1

0.3

0.5

0.7 Quantity of Information

Computationvalue p(value) p(poor | value) QI (value) p(value) * QI (value)[0,10[ 0.051 0.999 0.00924 0.000474

[10,20[ 0.25 0.938 0.232 0.0570323[20,30[ 0.26 0.732 0.581 0.153715

Decision Trees (4)

Limitations

I XOR-like attributes

I Attributes with many values

I Numerical attributes

I Overfitting

Limitations

Numerical Attributes

I Order the values val1 < . . . < valtI Compute QI([att < vali ])

I QI(att) = maxi QI([att < vali ])

The XOR caseBias the distribution of the examples

Complexity

Quantity of information of an attribute

n ln n

Adding a node

D × n ln n

Tackling Overfitting

Penalize the selection of an already used variable

I Limits the tree depth.

Do not split subsets below a given minimal size

I Limits the tree depth.

Pruning

I Each leaf, one conjunction;

I Generalization by pruning litterals;

I Greedy optimization, QI criterion.

Decision Trees, Summary

Still around after all these years

I Robust against noise and irrelevant attributes

I Good results, both in quality and complexity

Random Forests Breiman 00

Overview

Examples


Decision trees


Validation issues

1. What is the result ?

2. My results look good. Are they ?

3. Does my system outperform yours ?

4. How to set up my system ?

Validation: Three questions

Define a good indicator of quality

I Misclassification cost

I Area under the ROC curve

Computing an estimate thereof

I Validation set

I Cross-Validation

I Leave one out

I Bootstrap

Compare estimates: Tests and confidence levels

Which indicator, which estimate: depends.

Settings

I Large/few data

Data distribution

I Dependent/independent examples

I balanced/imbalanced classes

Overview

Examples


Decision trees


Performance indicators

Binary class

I h∗ the truth

I h the learned hypothesis

Confusion matrix

h / h∗ 1 0

1 a b a+b0 c d c+d

a+c b+d a + b + c + d

Performance indicators, 2

h / h∗ 1 0

1 a b a+b0 c d c+d

a+c b+d a + b + c + d

I Misclassification rate b+ca+b+c+d

I Sensitivity (recall), True positive rate (TP) aa+c

I Specificity, False negative rate (FN) bb+d

I Precision aa+b

Note: always compare to random guessing / baseline alg.

Performance indicators, 3

The Area under the ROC curve

I ROC: Receiver Operating Characteristics

I Origin: Signal Processing, Medicine

Principle

h : X 7→ IR h(x) measures the risk of patient x

h leads to order the examples:+ + +−+−+ + + +−−−+−−−+−−−−−−−−−−−−

Given a threshold θ, h yields a classifier: Yes iff h(x) > θ.+ + +−+−+ + ++ | − − −+−−−+−−−−−−−−−−−−

Here, TP (θ)= .8; FN (θ) = .1

The ROC curve

θ 7→ IR2 : M(θ) = (1− TNR,FPR)

Ideal classifier: (0 False negative,1 True positive)Diagonal (True Positive = False negative) ≡ nothing learned.

ROC Curve, Properties

PropertiesROC depicts the trade-off True Positive / False Negative.

Standard: misclassification cost (Domingos, KDD 99)

Error = # false positive + c × # false negative

In a multi-objective perspective, ROC = Pareto front.

Best solution: intersection of Pareto front with ∆(−c ,−1)

ROC Curve, Properties, foll’dUsed to compare learners Bradley 97

multi-objective-likeinsensitive to imbalanced distributionsshows sensitivity to error cost.

Area Under the ROC Curve

Often used to select a learnerDon’t ever do this ! Hand, 09

Sometimes used as learning criterion Mann Whitney

Wilcoxon

AUC = Pr(h(x) > h(x ′)|y > y ′)

WHY Rosset, 04

I More stable O(n2) vs O(n)

I With a probabilistic interpretation Clemencon et al. 08

HOW

I SVM-Ranking Joachims 05; Usunier et al. 08, 09

I Stochastic optimization

Overview

Examples


Decision trees


Validation, principle

Desired: performance on further instances

Further examples

WORLD

h

Quality

Dataset

Assumption: Dataset is to World, like Training set is to Dataset.

Training set

h

Quality

Test examples

DATASET

Validation, 2

Training set

hTest examples Learning parameters

DATASET

perf(h)

Unbiased Assessment of Learning Algorithms

T. Scheffer and R. Herbrich, 97

Validation, 2

Training set


DATASET

parameter*, h*, perf (h*)

perf(h)



Validation, 2

Training set


DATASET

Validation set

True performance

parameter*, h*, perf (h*)

perf(h)



Overview

Examples


Decision trees


Confidence intervalsDefinitionGiven a random variable X on IR, a p%-confidence interval isI ⊂ IR such that

Pr(X ∈ I ) > p

Binary variable with probability εProbability of r events out of n trials:

Pn(r) =n!

r !(n − r)!εr (1− ε)n−r

I Mean: nε

I Variance: σ2 = nε(1− ε)Gaussian approximation

P(x) =1√

2πσ2exp−

12x−µσ

2

Confidence intervals

Bounds on (true value, empirical value) for n trials, n > 30

Pr(|xn − x∗| > 1.96√

xn.(1−xn)n ) < .05

z ε

Tablez .67 1. 1.28 1.64 1.96 2.33 2.58ε 50 32 20 10 5 2 1

Empirical estimates

When data abound (MNIST)

Training Test Validation

Cross validationFold

2 31

Run

N

2

1

N

Error = Average (error on

N−fold Cross Validation

of h

learned from )

Empirical estimates, foll’d

Cross validation → Leave one out

2 31

Run 2

1

Fold

n

n

Leave one out

Same as N-fold CV, with N = number of examples.

PropertiesLow bias; high variance; underestimate error if data notindependent

Empirical estimates, foll’d

Bootstrap

Dataset

Training set

Test set.

rest of examples

with replacement

uniform sampling

Average indicator over all (Training set, Test set) samplings.

Beware

Multiple hypothesis testing

I If you test many hypotheses on the same dataset

I one of them will appear confidently true...

More

I Tutorial slides:http://www.lri.fr/ sebag/Slides/Validation Tutorial 11.pdf

I Video and slides (soon): ICML 2012, Videolectures, TutorialJapkowicz & Shahhttp://www.mohakshah.com/tutorials/icml2012/

Validation, summary

What is the performance criterion

I Cost function

I Account for class imbalance

I Account for data correlations

Assessing a result

I Compute confidence intervals

I Consider baselines

I Use a validation set

If the result looks too good, don’t believe it

Master Recherche IAC TC2: Apprentissage Statistique & …sebag/Slides/M2R_TC2_1_cours.pdf · Master...

Documents

Transcript of Master Recherche IAC TC2: Apprentissage Statistique & …sebag/Slides/M2R_TC2_1_cours.pdf · Master...